Integrating Item Hierarchy Relationship And User Preference For Recommendation

With the popularization of the Internet and the development of information technology,while enjoying various convenient network services,people also suffer from the inability to distinguish and make choices in the face of massive information.As an effective tool to alleviate information overload,recommender systems have been deployed in major network platforms to improve user experience and increase merchants' revenue.In recent years,they have made great progress in academia and industry.In this paper,the recommendation for e-commerce scenarios is used to improve the click-through rate of the recommendation models and alleviate the data sparsity of the target behavior.The main work is as follows:(1)A series of user behaviors on e-commerce websites often imply rich user preference information,and these behaviors have certain uncertainty due to the diversity and suddenness of user interests,and there is also uncertainty in the process of user interaction with the next item.Deterministic attention mechanism reflect the randomness of the model in the output layer,but a single output layer cannot effectively model the uncertainty in the user's historical behavior sequence.And because it regards each user behavior as an independent individual,it cannot capture the hierarchical dependencies among user behaviors.In response to the above problems,this paper applies Bayesian attention to recommender systems,and proposes a Bayesian Attention-Based Recommender System(BARS).Specifically,this paper regards the attention weights as data-dependent local random variables,and learns the attention distribution by approximating its posterior distribution under the Bayesian framework,and introduces a prior distribution with contextual information that depends on the user's historical behavior sequence to provide empirical uncertainty for the learning of the attention distribution to control the variance.For the sequential relationship between user behaviors,this paper applies Bi-LSTM to model,so that the current user's interest has past and future information.Experimental results on Advertising and Amazon datasets demonstrate the effectiveness of the proposed model.(2)Users have multiple interests at the same time,and different user behaviors reflect different user preferences for items.Aiming at the problem of effectively exploiting user multi-behavior data,this paper proposes a multi-behavior model based on attention network and transfer(MBATT).In this paper,user behaviors are sorted according to the shopping process,followed by browsing,adding to shopping cart,and purchasing.As the shopping process progresses,the user pays an increasing of cost,which makes the interaction data more and more sparse.Therefore,the pre-order behaviors as auxiliary behaviors can alleviate the data sparsity of the post-order target behavior.However,the pre-order behaviors contains a lot of noise due to user's mis-operation,so this paper shares the user embedding and item embedding in all behaviors,so that the subsequent behavior can provide guidance for the pre-order behaviors and subsequent behavior help to denoise for pre-order behaviors to a certain extent.In a single behavior,this paper uses the attention network to focus on the items related to the user's interest,and exploit the conscious user's intention in the unconscious user interaction set,so that the compressed item vector contains enough information to represent the user's interest.Then apply matrix factorization to predict the possibility of user interaction with the next item,and utilize transfer to realize the transfer between user behaviors.The experimental results on Beibei and Taobao datasets demonstrate the rationality of our model.